VeriPhy: Agentic Physical Reasoning for World Model Evaluation and Refinement

arXiv 2026 30.4 method

TLDR

VeriPhy is an auditable physical-verification system that evaluates generated videos for physical obligations, outperforming prior evaluators on a human-annotated corpus.

Reasoning

The paper presents a novel provenance-based evaluation method and benchmark for physical reliability in generated video, with strong auditable evidence. The main weakness is that the abstract is truncated, and the system depends on frozen expert modules and an external backbone, limiting insight into end-to-end refinement.

Read-first score

Read-first score 30.4, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 31.

Recency 5%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Methodology quality 15%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=evaluation,result

Topical relevance 25%

44.3

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Reproducibility 15%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Graph prestige 15%

5.7

PageRank-style prestige over the local citation graph built from available OpenAlex references. local_pagerank=0.0016

Citation impact 15%

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. cited_by_count=0

Citation velocity 10%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 207.

Keyword Scores

world model

video world model

generative world model

world dynamics prediction

world simulator

interactive world model

model-based reinforcement learning world model

Deep Analysis

Innovations

VeriPhy introduces an auditable physical-verification system that compiles prompts into typed physical obligations and a statically validated execution plan before any frames are observed.
During execution, observations gate and scope only declared calls to frozen low-level experts such as segmentation/tracking, counting, eleven typed physical measurements, depth, OCR, and audio-event detection.
Each expert action returns a provenance-carrying evidence record whose payload is either a typed measurement or an explicitly tagged learned state.
Typed resolvers and fixed composition map usable evidence records to a three-valued state (supported, contradicted, or unknown; surfaced as plausible, implausible, or abstain) with full provenance.
Evaluation is anchored in a 1,500-clip corpus of human-annotated flaw records that localize real generation failures in prompt reference, space, and time.
Each verdict retains its evidence record and provenance, making traces auditable per decision and usable as an interface for writing critic verdicts back into generation.

Methodology

VeriPhy uses a text-only planner to compile the prompt into typed physical obligations and a statically validated execution plan before observing frames. During execution, only declared calls to frozen low-level experts are gated and scoped by observations, and each call returns a provenance-carrying evidence record. Typed resolvers and fixed composition then map usable records to supported, contradicted, or unknown verdicts with full provenance.

Key Results

On a 149-clip core carrying 304 human-annotated flaw records, VeriPhy accounts for 228 records, compared with 164 for a published question-decomposition evaluator given the same clips and claims. Prompting the same backbone monolithically reaches 222, so the main separation is auditable provenance rather than recall alone.

Resources

Paper